Reflection AI 开源 Beam：501B 参数 MoE，号称最强西方开源模型
Reflection AI 开源了 Beam，501B 参数 MoE，作者称推理算力只要 GLM-5.2 的三分之一，Apache 2.0 可商用，不过注意这数字只是估算。
Reflection AI 发布开源模型 Beam，为稀疏 MoE 架构，总参数 501B、激活 23B，在 6,144 块 GB300 GPU 上用不到 4 周预训练了 23.8T tokens，支持 1M token 有效上下文。RL 阶段使用 10.5K 块 GB300 训练 4 周，产生超 1 亿次 rollouts。Beam 在 Terminal Bench v2.1 上得 80.1，接近 GLM-5.2 的 81.0，但落后于 DeepSeek V4.1 Flash 的 90.6 和 Kimi K3 的 88.3。所称 3-4 倍于 GLM-5.2 的推理效率优势基于前向 FLOPs 估算，未计入 prefill、attention 和服务开销。模型将以 Apache 2.0 协议发布 FP8 和 NVFP4 版本。
The much anticipated open-source model from Reflection AI just dropped.
Beam now becomes the strongest western open model, rivals GLM-5.2 while using 3-4x less inference compute.
> The 3-4x efficiency edge over GLM-5.2 rests on estimated forward-pass FLOPs (2 × active parameters × generated tokens), which leave out prefill, attention and serving overhead, so it is not a measured cost.
> A sparse mixture-of-experts model with 501B total and 23B active parameters, pretrained on 23.8T tokens in under 4 weeks on 6,144 GB300 GPUs, with a 1M-token effective context.
> The RL run used 10.5K GB300 GPUs for 4 weeks, produced over 100M rollouts across nearly 1M environments and about 1.3B sandboxes, and ended with no sign of a plateau.
> Beam nearly ties GLM-5.2 at 80.1 versus 81.0 on Terminal Bench v2.1, but trails DeepSeek V4.1 Flash at 90.6 and Kimi K3 at 88.3, neither of which appears in the headline chart.
and Reflection will also ship FP8 and NVFP4 builds under Apache 2.0.
